AI Models

DeepSeek V4 Flash, Running in MoClaw

DeepSeek V4 Flash is what MoClaw's Fast tier runs: 1M context, agent-tuned, about a third of V4 Pro's price. Pick Fast, type a task, go.

DeepSeek V4 Flash

MoClaw is an independent managed AI computer where you can run frontier models with no setup or self-hosting.

All integrations

See DeepSeek V4 Flash in action

Community demos and open-source work built on DeepSeek V4 Flash, not MoClaw output.

DeepSeek V4 Flash demo — A stock-report agent that writes its own .docx, built by @farhanhelmycode

A stock-report agent that writes its own .docx

DeepSeek V4 Flash driving a web agent through 12 steps and 31 tool calls in 11.8 seconds, ending in a finished Word report built by Python inside a Docker sandbox. The kind of long tool chain Flash is cheap enough to run repeatedly.

Demo by @farhanhelmycode Open
DeepSeek V4 Flash demo — Running V4 Flash locally on Metal, CUDA and ROCm, built by antirez

Running V4 Flash locally on Metal, CUDA and ROCm

ds4 is a 21.4k-star MIT-licensed inference engine written in C for running DeepSeek V4 Flash and V4 Pro on your own hardware, with DSpark speculative decoding. Evidence of how much work the self-hosted route actually is.

Open source · antirez/ds4 Open

How it works

3 steps to wire up DeepSeek V4 Flash, no engineering required.

  1. 1

    Pick Fast in the model picker

    Fast is DeepSeek V4 Flash. There's no key to paste and no DeepSeek account to create; MoClaw handles the provider side.

  2. 2

    Give it a task, not a prompt

    Flash runs on a real computer here, so ask for the outcome. Read these files, check that site, put the result in a spreadsheet.

  3. 3

    Move up a tier when it stalls

    If a task needs more reasoning, switch tiers mid-thread. Files, history, and context carry over.

Why DeepSeek V4 Flash Is MoClaw's Fast Tier

What is DeepSeek V4 Flash?

DeepSeek V4 Flash is the smaller, faster half of the V4 family, and it went GA on July 31, 2026 as build DeepSeek-V4-Flash-0731. DeepSeek kept the preview's architecture and size and re-ran post-training, which is why the capability jump surprised people: on DeepSeek's own numbers the GA Flash beats the V4 Pro preview across the agent benchmarks it published.

It carries the same 1-million-token context window and 384K output ceiling as V4 Pro, the same thinking and non-thinking modes, tool calling, JSON output, and the same OpenAI- and Anthropic-compatible endpoints. Like the rest of the family, it has no vision input.

DeepSeek V4 Flash benchmarks

DeepSeek published Terminal Bench 2.1 at 82.7, DeepSWE at 54.4, NL2Repo at 54.2, CyberGym at 76.7, and Toolathlon-Verified at 70.3. For a model at this price that is an odd set of numbers to be looking at, and the developer reaction reflected it: on Hacker News, more than one person described Flash 0731 as the bigger leap of the two releases, and V4 Pro's GA a fortnight later as the quieter one.

All of it is vendor-reported. DeepSeek did open-source the agent framework it benchmarked with, DeepSeek Harness, under MIT as a developer preview, so the scaffolding is at least public.

DeepSeek V4 Flash pricing

As published, Flash is $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output. From 16:00 UTC on August 16, 2026 the V4 family switches to peak and off-peak billing: Flash lands at $0.44 in and $1.32 out at peak, and $0.22 and $0.66 off-peak, with peak defined as 01:00 to 04:00 and 06:00 to 10:00 UTC.

The concurrency limit is 2,500 against V4 Pro's 500, which matters more than the price for anything batch-shaped.

Running DeepSeek V4 Flash on MoClaw

On MoClaw, Flash is the Fast tier. Pick Fast in the model picker and you are running DeepSeek V4 Flash on a managed cloud computer: a persistent file system, a browser that keeps its session, a shell, scheduling, and 50-plus skills already connected. DeepSeek runs platform-managed here, so there's no DeepSeek account to open, no key to store, and no peak-hour clock to reason about. Plan credits cover it.

Switching is one click. Start something on Fast, decide it needs more depth, and move the same thread to a heavier tier without losing your files or history. That is the shape most people actually work in: cheap model for the volume, expensive model for the hard part.

Where Flash stops making sense

It isn't the model for a task where being wrong is expensive and you won't be checking. Developers who use DeepSeek heavily tend to say the same thing: it earns its price when you steer it, and they don't hand it fully autonomous work they can't verify. Anything image-based is out entirely. For long-horizon autonomy on a codebase you care about, reach for a heavier tier and let Flash do the parts you can check at a glance.

DeepSeek V4 Flash on MoClaw vs the DeepSeek API

Pick the one that fits how you use AI.

MoClaw (Fast tier) DeepSeek API direct
Setup Pick Fast. That's the whole setup. Account, API key, prepaid balance, and your own client code.
What it can touch Browser, files, shell, schedules, and 50-plus skills. Text in, text out. Tools are yours to build.
Billing Plan credits, no peak-hour rates to plan around. Per token, split into peak and off-peak from August 16.
Moving to a stronger model Switch tier mid-thread, keep the files and history. Different provider, different endpoint, different bill.

FAQ

Quick answers about pricing, privacy, and limits.

What is DeepSeek V4 Flash?
The smaller model in DeepSeek's V4 family, GA since July 31, 2026 as build DeepSeek-V4-Flash-0731. It keeps the 1M-token context window and 384K output ceiling of V4 Pro, supports thinking and non-thinking modes and tool calling, and costs roughly a third as much.
Can I run DeepSeek V4 Flash on MoClaw?
Yes. Flash is the model behind MoClaw's Fast tier. Pick Fast in the model picker and DeepSeek runs on your MoClaw cloud computer with a browser, files, a shell, and scheduling already connected. No DeepSeek account and no API key.
How much does DeepSeek V4 Flash cost?
On DeepSeek's API it's $0.14 per million input tokens and $0.28 per million output, moving on August 16, 2026 to $0.22/$0.66 off-peak and $0.44/$1.32 at peak. On MoClaw it runs on plan credits, so there's no per-token metering and no peak-hour clock.
Is V4 Flash good enough, or should I use V4 Pro?
For volume work with a clear plan, Flash holds up: 82.7 on Terminal Bench 2.1 against Pro's 87.9. Pro pulls ahead on the long autonomous runs, particularly on DeepSWE where it scores 62.7 against Flash's 54.4. A common pattern is planning on a heavy model and executing on Flash.
Does DeepSeek V4 Flash handle images?
No. The V4 family has no vision input, so screenshots and scanned documents need a different model for that step.
Do I need a DeepSeek API key for the Fast tier?
No. DeepSeek runs as a platform-managed route inside MoClaw. There's no key to paste, no balance to top up, and no separate provider invoice.
What is the context window on DeepSeek V4 Flash?
One million tokens in and up to 384K out, the same as V4 Pro. Long transcripts, whole repositories, and large exports fit in a single pass.
Can I switch away from Flash mid-conversation?
Yes. Move the thread to a stronger tier whenever a task needs more reasoning; your files, history, and context follow. Most people run cheap for volume and switch up for the hard part.

Try MoClaw free.

Start with 1,000 credits free.

Cancel anytime